3 messages in org.freebsd.freebsd-auditRe: Need reviewers for busdma changes...| From | Sent On | Attachments |
|---|---|---|
| Matthew Jacob | Aug 10, 2001 3:44 pm | |
| Bill Paul | Aug 10, 2001 4:39 pm | |
| Matthew Jacob | Aug 10, 2001 4:43 pm |
| Subject: | Re: Need reviewers for busdma changes to ethernet driver | |
|---|---|---|
| From: | Bill Paul (wpa...@FreeBSD.ORG) | |
| Date: | Aug 10, 2001 4:39:40 pm | |
| List: | org.freebsd.freebsd-audit | |
Seems mostly good.
1. Nit:
/* * Obtain the physical addresses of the RX and TX * rings which we'll need later in the init routine. */
More formally, it's not a 'physical address'- it's a "Bus Address", or "Address appropriate for a device on this bus to use to access the memory so mapped". Under sparc64, for example, it would be an iommu mapping address.
2. All callback functions for dma- you should be checking error.
Ok...
3. You might want to recycle rather than create/destroy mbuf tags.
I'm not creating/destroying tags, I'm creating/destroying dma maps (bus_dmamap_t).
My understanding is that I need a dmamap for every buffer that I need to send to the NIC, and in the transmit case, I might have several buffers per packet (i.e. an mbuf list with all the packet fragments). But I can't know ahead of time how many fragments will be in the chain, so I can't really say "allocate this many dma maps for packet transmit."
You should note that FreeBSD's bus_dma is hardly as strict or as formal as NetBSD's, so there's room for a lot of slop here. The key thing to test it on is an alpha that has both direct-mapped and sgmap hardware to see that it works.
Hopefully I will be able to do this on Monday. In the meantime, I have also converted the RealTek driver:
http://www.freebsd.org/~wpaul/RealTek/busdma
This was somewhat easier because there's just one giant DMA buffer for receive and no descriptor structures.
-Bill
To Unsubscribe: send mail to majo...@FreeBSD.org with "unsubscribe freebsd-audit" in the body of the message